search

vuvko / fitw2020

Scripts for Families in the Wild 2020 challenge, track 1
MIT License
9 stars 3 forks source link

loss function #2

Closed garynlfd closed 2 years ago

garynlfd commented 2 years ago

Hi,

I am interested in RFIW and came across your paper. You mentioned that you replaced SphereFace with ArcFace. But in the paper, I think the loss function is more like a softmax loss function. I also viewed your code. In train.py, there is a variable named all_losses and you wrote three kinds of loss but finally used 'softmax'. Did you use 'arc' in RFIW? If I misunderstand something, please correct me. I would be grateful if you could give me some hints or feedbacks

Thanks!

vuvko commented 2 years ago

I haven't used margin-based losses for training. ArcFace is mentioned as a base model for finetuning and as a baseline.

garynlfd commented 2 years ago

Okay, I got it! Thank you so much!